Robotics · Daily digest
VLA Daily: What’s New in Robot Foundation Models
Every morning: new vision-language-action papers, model releases, deployments and funding, ranked for people who build VLAs and measured against the October 2026 state-of-the-art report. Every item carries a source tag.
PRIMARY read the paper or official page · SEARCH search snippets only · VENDOR company claim · SEC press coverage. Numbers without a source are left out.
VLA DailyMon Nov 2, 2026
A quiet three days at the big labs: no confirmed releases from Physical Intelligence, NVIDIA, Gemini Robotics, Figure, Skild or Generalist. The new ideas are two papers that add future-prediction supervision to the VLA recipe, and the money is flowing to human-demonstration data. Paper coverage was thin because arXiv and Hugging Face Papers could not be read directly.
New in the last 3 days
JAM (Just Action Model) and PDRVLA SEARCH
JAM argues that VLAs and world-action models are one framework: a pretrained visual backbone plus an action expert, trained on both action and foresight losses. PDRVLA adds a module that predicts future RGB, depth and optical-flow features, supervised by a frozen video world model.
Why it matters: Both add future-prediction supervision on top of the default VLM-plus-flow-expert recipe.
GenRobot AI raises USD 100M Series A+ SEC
The Beijing company records people doing physical tasks to produce robot training data; the round closed in Q3. Investors include SAIC Motor, Baidu and Galaxy General Robotics.
Why it matters: The money is going to human-demonstration data, not only to model builders.
Meta reportedly built a humanoid prototype with Fourier SEC
The prototype, reportedly called "Max", took about three months and is a research platform aimed at household robots, with a second version in progress. It could not be corroborated in a separate search.
Why it matters: Meta entering humanoids is outside the baseline report.
Catching up
Natura-1 (Nature Will) VENDOR
A 50B-parameter mixture-of-experts model for 20-DoF dexterous hands, pretrained on over 1M hours of first-person human video, with a claimed 100 ms control loop. It is compared only against π0.5 and a coding-based policy.
Why it matters: A much larger model and a human-video scale claim, with no independent evaluation yet.
Unitree UnifoLM-WLA-1.0 VENDOR
A roughly 6B humanoid VLA trained on about 2,500 h of real-robot data, running 64 tasks on the G1 with one set of weights. Several outlets call it open source, but as of Sep 15 only UnifoLM-ER-1 (4B) and UnifoLM-ER-Flow were public; the WLA-Base weights were not.
Why it matters: A whole-body humanoid VLA, but the weights are not confirmed downloadable.
Other robotics funding SEC
Bedrock Robotics reportedly raised USD 600M at a USD 4.1B valuation (date unclear). Foundry Robotics raised USD 100M for an autonomous assembly factory, and Bleu Robotics raised EUR 5M pre-seed.
Why it matters: Capital is also flowing to construction autonomy and factory automation.
Watch
- Whether Unitree releases the WLA-Base weights, and under what licence.
- Whether JAM- and PDRVLA-style future-prediction heads show up in large-lab models.
- Meta's humanoid effort, and whether a VLA or world-model release comes with it.
Not verified: search snippets mentioned GR00T N2, Cosmos 3 and LeRobot v0.6.0, but only from low-quality sources, so they are left out.